Rombo Denoiser

Here we use a non-local sample-based approach together with feature-based filtering and adaptive sampling. A proof of concept on latest denoising tech plus some creative engineering that makes it working consistently for production.

For example, with a little help from the camera and with Arnold in kinda ‘pathtrace’ mode (AA samples 4×4 while all others set to 1) works incredibly good on DOF and highlights !

metal_params_normals

Here we can really see how strong HDR highlights come out well denoised while keeping all the HDR range and IBL detail. Look how well the little logo is preserved, wheels rims etc. Denoising time is really cheap, here on a full HD image it takes just 10 seconds. Click the image to see it full resolution.

metal_params_normals

Testing textures, penumbra and strong highlights … look how well details on the pillow remains vivid with no blur or artifacts. Reconstruction on the aliased lamp border looks pretty natural and dark areas are not blotchy !

metal_params_normals

A Cornell Box testing how refractions get denoised.

metal_params_normals

Here some comparisons to see that for strong highlights we get better result with a 4x4AA + denoiser then a full 10x10AA.

metal_params_normals

Final denoised image. Notice how ground texture is still super sharp after denoising took place.

metal_params_normals

Works really well also on moblurred scenes.

metal_params_normals

And on global illumination. Look also how artifacts on the windows are filtered out.

metal_params_normals

Eventually the denoiser implementation as an Arnold driver via ASS files.

Leave a Reply